Assembling the brief…

Skip to content

Preprints

Preprints, newest first.

Not peer reviewed. Posted to a preprint server and read on that basis.

Preprints

Preprint · not peer reviewedUpdated Sep 2 · revised preprint

Characterizing the landscape of gene process dependencies in cancer

A revised BioBombe preprint uses machine learning to identify groups of genes that cancer cells depend on, then links those dependencies to drug sensitivity and tests a cladribine prediction in pediatric glioma cells. The approach could broaden target discovery from individual genes to vulnerable biological processes, with new RNA-seq modelling and wet-lab validation added in this update. Killing cultured cancer cells does not establish clinical benefit or justify using cladribine to treat these patients.

First author: Curd, J. B.Last author: Way, G. P.

Affiliations: University of Colorado Anschutz

Source: bioRxiv / medRxiv

cancer biology

Preprint · not peer reviewedSep 2

BanffNET, a Deep Learning System for Comprehensive Histological Lesion Quantification in Kidney Transplant Biopsies

BanffNET uses deep learning to score kidney-transplant biopsy slides, with training on 7,249 slides and external testing on 11,028 slides across five datasets. Its reported agreement with expert assessments and molecular signs of rejection suggests a route to more consistent, scalable measurement of transplant injury. This preprint does not establish that deploying the system improves transplant treatment decisions or patient outcomes.

First author: Giorgio BuzzancaLast author: Jesper Kers

Affiliations: Department of Pathology, Systems Pathology Research Group, Leiden University Medical Center, Leiden University, Leiden, Netherlands

Source: bioRxiv / medRxiv

More: Europe PMC

pathology

Preprint · not peer reviewedSep 2

Can Dental AI Really Beat Dentists? DentalPair-Cert for Rigorous AI-Dentist Inference

DentalPair-Cert shows why separately reported accuracy scores can leave comparisons between dental AI and dentists with unresolved statistical uncertainty. The authors propose reporting how often both get the same case right, allowing reviewers to reconstruct a properly paired comparison instead of assuming independence. This preprint offers statistical methods and simulations, not evidence that a particular dental AI improves patient care.

First author: Alve, S. R.Last author: Meem, S. M. A. C.

Affiliations: The University of Texas at Dallas

Source: bioRxiv / medRxiv

dentistry and oral medicine

Preprint · not peer reviewedSep 2

Default-filled outcome labels in a deployed cognitive-screening programme: an operator-level audit and the construction of twenty-four language-model arms

An audit of a cognitive-screening programme found that accounts recording no impairments contributed 40.5% of routine outcome labels, highlighting a serious training-data problem. Comparisons of 24 modelling approaches showed why auditing labels and testing against a simple regression baseline can matter more than adding model complexity; a distilled language model improved discrimination in this dataset. The preprint's specialist-labelled cohort was small, so these results do not establish a general recipe for clinical deployment.

First author: Jun JiLast author: Huali Wang

Affiliations: Peking University Sixth Hospital, Peking University Institute of Mental Health, NHC Key Laboratory of Mental Health (Peking University), National Clinical Research Center for Mental Disorders (Peking University Sixth Hospital)

Source: bioRxiv / medRxiv

health informatics

Preprint · not peer reviewedUpdated Sep 2 · first posted Jan 7, 2026

Accessible and reproducible deployment reveals the practical boundaries of single-cell foundation models

An expanded evaluation of 13 single-cell foundation models across nearly 100 datasets found that larger models did not reliably outperform simpler methods, with the clearest gains appearing when labelled data were extremely scarce. The study offers practical guidance on when expensive pretraining helps and when a conventional method may be sufficient, including limits in spatial and regulatory tasks. This is a revised preprint, and its conclusions remain dependent on the evaluated datasets, tasks and comparison settings.

First author: Hou, S.Last author: Zhou, X.

Affiliations: Yale University

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedSep 1

From Public Archive to Reusable Resource: Characterizing Gut Microbiome Metadata in the NCBI SRA

An audit of public gut-microbiome sequencing records found abundant technical metadata but inconsistent biological context; host identity could be assigned to only 13% of samples in the broadly labelled gut-metagenome cohort. The findings expose a data-preparation bottleneck for reproducible cross-study analysis and microbiome foundation models: a large archive is not automatically a usable training set. This preprint evaluates metadata quality, not the performance of a new biological prediction model.

First author: Oderman, K. A.Last author: Hampton-Marcell, J. T.

Affiliations: University of Illinois Chicago

Source: bioRxiv / medRxiv

More: Europe PMC

bioinformatics

Preprint · not peer reviewedUpdated Sep 1 · first posted Dec 22, 2025

Spatial Transcriptomics As Rasterized Image Tensors (STARIT) characterizes cell states with subcellular molecular heterogeneity

An expanded STARIT preprint turns spatial transcript measurements inside cells into image tensors, letting vision models identify cell-state differences that ordinary gene-count summaries miss. The new version adds simulation tests and feature-extractor benchmarks, strengthening a practical route to studying subcellular organization rather than expression levels alone. The work remains a preprint, and results depend on the imaging data and choice of vision features.

First author: Velazquez, D.Last author: Fan, J.

Affiliations: Johns Hopkins University

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedSep 1

Scaling recipes for single-cell RNA sequencing foundation models: when do scaling laws hold?

Researchers studying single-cell foundation models found that pretraining loss followed scaling relationships with model size and compute, and derived guidance for learning rates and model proportions. These results could help teams budget training runs and choose architectures for transcriptomic models instead of relying solely on trial and error. Lower pretraining loss is not itself evidence of better biological predictions or discovery outcomes.

First author: Borra, F.Last author: Buffa, F. M.

Affiliations: University of Oxford

Source: bioRxiv / medRxiv

More: Europe PMC

bioinformatics

Preprint · not peer reviewedSep 1

TigerAI: An AI-powered genetic evidence platform to support clinical development

The TigerAI team evaluated GPT-5-generated genetic evidence across 13,022 target–indication pairs and reported that it was at least as informative as expert-curated evidence for inferring clinical success. The study suggests a route to scaling genetic support for drug-target decisions, with an accompanying platform for querying evidence and comparing models. The preprint does not show that using the system prospectively improves clinical-trial success.

First author: Zhang, Z.Last author: Zhao, B.

Affiliations: University of Pennsylvania

Source: bioRxiv / medRxiv

genetics

Preprint · not peer reviewedUpdated Aug 31 · first posted May 26, 2026

Prioritizing peptides for targeted mass spectrometry experiments using deep learning

An expanded Bromo preprint uses a transformer trained on peptide-pair measurements to rank targets for mass-spectrometry assays, accounting for charge state and adaptation to local experimental conditions. Better peptide selection could make quantitative proteomics more reliable across sample preparations and instruments. The reported advantages are dataset comparisons, not a guarantee of performance in every laboratory.

First author: Sonthalia, S.Last author: Noble, W. S.

Affiliations: University of Washington

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedUpdated Aug 31 · first posted Mar 17, 2026

Spatial agent-based modeling and interpretable machine learning identify determinants of combination-therapy response in HER2-heterogeneous breast cancer

An expanded breast-cancer modelling preprint combines spatial tumor simulations with an interpretable random forest to examine how mixed HER2 states influence treatment escape. Controlled spatial comparisons suggest that where resistant cell states sit can matter even when tumor size and cell counts match. The combination-treatment results are simulations, not evidence that the proposed regimen benefits patients.

First author: Rahman, N.Last author: Jackson, T. L.

Affiliations: University of Michigan

Source: bioRxiv / medRxiv

cancer biology

Preprint · not peer reviewedAug 31

A Pragmatic Randomized Trial of an EHR-Integrated Generative AI Chart Summarization Tool for Ambulatory Clinicians

In a 90-day randomized study of 284 outpatient clinicians, an AI chart-summary tool reduced reported task load but did not significantly reduce charting time, and engagement declined. The findings show why evaluating clinical AI requires measuring actual use and workflow effects, not just the quality of generated text. This single-system preprint also reports concerns about inaccuracies, so the absence of reported safety events should not be read as proof of safety.

First author: Chin, A. T.Last author: Lukac, P. J.

Affiliations: UCLA Health Information Technology, UCLA Health, University of California, Los Angeles, Los Angeles, CA, USA; Department of Pediatrics, Division of Immunology,

Source: bioRxiv / medRxiv

health informatics

Preprint · not peer reviewedAug 28

PathFold: Predicting the Entire Protein Folding Pathway from Protein Sequence Alone

PathFold combines an AlphaFold-based component with a diffusion model to generate a proposed folding trajectory directly from a protein sequence, including intermediate structures. Moving beyond a single static structure could help researchers investigate how proteins reach their final shape and where folding goes wrong. The reported agreement with simulations and experimental folding measurements does not establish accurate physical kinetics for arbitrary proteins.

First author: Zhang, Z.Last author: Kihara, D.

Affiliations: Purdue University

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedAug 28

The first OpenBind release: An open experimental structure-affinity dataset and benchmark for structure-based AI

The first OpenBind release pairs experimental structures and affinity measurements for an enteroviral protease, including 925 crystallographic binding events and affinity data for 601 compounds. A coherent experimental dataset gives drug-discovery AI a harder test than disconnected structures and exposes weaknesses in pose ranking, affinity prediction and virtual screening. The benchmark is centred on one target, so improvements after target-specific fine-tuning do not establish broad generalization.

First author: Nelen, J.Last author: Imrie, F.

Affiliations: University of Oxford

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedAug 26

A Mammalian High-Throughput Screen for AI-Designed Peptide-Guided Protein Degraders

Researchers paired protein-language-model peptide design with pooled human-cell screening to identify guides that direct targeted protein degradation across four targets. Testing designs inside mammalian cells provides a practical way to weed out candidates that look promising computationally but fail to change protein abundance. The preprint reports reporter and cell-line experiments, not evidence of benefit in animals or patients.

First author: Lin ZhaoLast author: Pranam Chatterjee

Affiliations: University of Pennsylvania, University of Wisconsin, Children's Hospital of Philadelphia

Source: bioRxiv preprint

Preprint
Protein design
Drug discovery

Preprint · not peer reviewedAug 26

Temporal dynamics improves machine learning-based prediction of cell state from quantitative phase imaging

A spatiotemporal neural network classified six cell states from label-free imaging more accurately as additional frames were added. Label-free time-series models could distinguish dynamic cell states that look similar in a single frame.

First author: Alizada, S.Last author: Zangle, T. A.

Affiliations: University of Utah

Source: bioRxiv preprint

Preprint
Data & benchmarks

Preprint · not peer reviewedAug 26

Assay concordance sets exact ceilings on what one biological score can predict

A model-free analysis argues that disagreement among biological assays places a measurable ceiling on predictor performance. Biological-AI benchmarks may misstate model headroom when they ignore disagreement among the experiments used as ground truth.

Author: Liu, Z.

Affiliations: Columbia University

Source: bioRxiv preprint

Preprint
Data & benchmarks

Preprint · not peer reviewedAug 26

Unobserved Sequence Space Has Many Functional Proteins

Wet-lab tests found many functional proteins far from known natural sequences while exposing inconsistent predictions from protein language models. Protein design has a larger experimental search space than nature has sampled, but computational scores still need laboratory validation.

First author: Diplock, N.Last author: Leonard, S. P.

Affiliations: Lawrence Livermore National Laboratory

Source: bioRxiv preprint

Preprint
Protein design
Data & benchmarks

Preprint · not peer reviewedAug 26

A mechanism-annotated benchmark reveals limited fidelity to drug-response signatures in single-cell perturbation models

A mechanism-aware benchmark found that high expression similarity often failed to preserve the drug-response signatures models were meant to predict. Drug-perturbation models need tests that reward biological mechanism, not merely reconstruction of average expression patterns.

First author: Li, L.Last author: Fang, B.

Affiliations: MindFlow.ai

Source: bioRxiv preprint

Preprint
Single-cell & spatial
Data & benchmarks
Drug discovery

Preprint · not peer reviewedAug 26

scDisent: regulatory-aware disentangled representation learning for multi-omic single-cell analysis

scDisent separates expression-linked and regulation-linked variables instead of compressing multi-omic measurements into one latent space. Separating regulatory state from expression could make single-cell embeddings more interpretable for perturbation analysis.

Author: Xi, G.

Affiliations: Jena University Hospital

Source: bioRxiv preprint

Preprint
Single-cell & spatial
Foundation models

Preprint · not peer reviewedAug 26

Assessing the translation of AI-prioritized genome-derived peptide fragments into validated antimicrobial candidates

Researchers mined yeast genomes, used AI to rank peptide candidates, and tested synthesized designs against Staphylococcus aureus, linking computational selection to measurable antimicrobial activity. The study tests whether biologically grounded sequence searches can produce useful experimental hits rather than just attractive model scores. The preprint also reports biases and generalization limits; laboratory activity does not establish a safe or effective antibiotic for patients.

First author: Ojeda, S.Last author: Munoz-Camargo, C.

Affiliations: Universidad de los Andes

Source: bioRxiv preprint

Preprint
Protein design
Drug discovery

Preprint · not peer reviewedAug 26

Robustness to nuisance perturbations enables unsupervised evaluation of single-cell foundation models

An unsupervised stress test separated single-cell foundation models that looked nearly equivalent on standard downstream benchmarks. Robustness tests can reveal fragile biological representations that polished embedding plots and task labels miss.

First author: Sallam, A.Last author: Gillis, J.

Affiliations: University of Toronto

Source: bioRxiv preprint

Preprint
Single-cell & spatial
Foundation models
Data & benchmarks

Preprint · not peer reviewedAug 26

Mapping Alzheimer's neuropathology signatures to the whole brain transcriptome using machine learning data fusion

A data-fusion model projected cortical Alzheimer's signatures across a 108-region whole-brain single-cell atlas. The approach offers a computational route around the cortical sampling bias that limits much Alzheimer's single-cell research. The whole-brain results are model-based extrapolations rather than direct pathology measurements in every region.

First author: Bhattacharya, A.Last author: Bzdok, D.

Affiliations: McGill University

Source: bioRxiv preprint

Preprint
Single-cell & spatial
Data & benchmarks

Preprint · not peer reviewedAug 26

A multi-scale structural and biophysical atlas of TCR-peptide-HLA recognition dynamics

DynaTPH adds standardized molecular-dynamics trajectories and biophysical features to 256 representative T-cell-receptor complexes. A standardized dynamics corpus could support more realistic computational immunology, TCR engineering, and model evaluation.

First author: Zhang, S.Last author: Fu, L.

Affiliations: Institute of Medical Artificial Intelligence, South China Hospital, Shenzhen University

Source: bioRxiv preprint

Preprint
Data & benchmarks

Preprint · not peer reviewedAug 22

A novel benchmark dataset for enzyme function prediction reveals the limitations of state-of-the-art models

EnzymARC tests whether enzyme-function predictors reject structure-guided decoys whose catalytic machinery has been disrupted. Benchmarks that include convincing negative examples can expose shortcut learning hidden by tests made only from annotated functional proteins. The decoys are expected to be inactive computationally; they were not experimentally validated as non-functional.

First author: Joao SartoriLast author: Lucas de Almeida Machado

Affiliations: Fiocruz, Instituto Oswaldo Cruz, Institute of Technology on Immunobiologicals (Bio-Manguinhos), Fiocruz

Source: bioRxiv preprint version 1

More: EnzymARC repository

Preprint
Data & benchmarks
Foundation models
Protein design

Preprint · not peer reviewedAug 20

bulk2scDiff: Generating Single-Cell Expression Profiles from Pseudobulk Data with Conditional Diffusion Models

A conditional diffusion model reconstructs plausible single-cell populations from bulk-like expression summaries. If it generalizes, the method could help explore cell-state hypotheses in studies that only have bulk expression data. Generated cells are hypotheses, not substitutes for measuring real single cells.

First author: Jiedan XiaoLast author: Andreas Raue

Affiliations: University of Augsburg

Source: bioRxiv preprint version 1

Preprint
Single-cell & spatial
Foundation models
Data & benchmarks

Preprint · not peer reviewedAug 20

A self-supervised DNA foundation model with collapse-resistant multimodal fusion

A DNA foundation model combines sequence with local and global chromatin accessibility while trying to prevent one modality from overwhelming the representation. Genomic foundation models may need cell-state context, not sequence alone, to represent regulatory function. The sole author discloses being founder and CEO of ModelsLive.

Author: Yuyan Chen

Affiliations: ModelsLive Inc.

Source: bioRxiv preprint version 1

Preprint
Foundation models
Single-cell & spatial
Data & benchmarks

Preprint · not peer reviewedAug 19

Supervised Learning of Phosphopeptide Sequence Constraints Enables Global Prediction of SH2 Domain Binding

SpY-C learns shared sequence constraints for whether a human phosphotyrosine site can participate in SH2-domain binding. A family-level filter could reduce the search space before researchers attempt the harder task of assigning a phosphosite to a particular SH2 domain. The classifier predicts general SH2-binding potential, not a specific domain–phosphosite interaction.

First author: Alekhya KandoorLast author: Kristen M. Naegle

Affiliations: University of Virginia, University of Southern Denmark, University of Connecticut

Source: bioRxiv preprint

More: SpY-C code

Preprint
Data & benchmarks
Drug discovery

Preprint · not peer reviewedAug 19

Single-cell foundation models benefit from cross-modal training: adding proteomics data beats parameter scaling

A small biological foundation model trained with proteomics matched or beat much larger RNA-only models on most reported benchmarks. Biological model progress may depend more on integrating complementary measurements than on scaling RNA-only architectures. All authors are employees and shareholders of Tesorai.

First author: Maximilien BurqLast author: Peter Cimermancic

Affiliations: Tesorai Inc.

Source: bioRxiv preprint version 1

Preprint
Foundation models
Single-cell & spatial
Data & benchmarks

Preprint · not peer reviewedAug 18

AlphaConformers: Structure-guided sampling enables prediction of multiple protein conformations

AlphaConformers uses structures from related proteins as template hypotheses to steer AlphaFold2 toward alternative conformations. Recovering more than one plausible protein conformation can reveal functionally relevant changes that a single predicted structure may conceal. The benchmark contains proteins with already known ligand-bound and unbound conformations.

First author: Julie Liliane DanielLast author: Diego Javier Zea

Affiliations: I2BC (CNRS, CEA, Université Paris-Saclay)

Source: bioRxiv preprint

More: AlphaConformers code

Preprint
Protein design
Foundation models
Data & benchmarks

Preprint · not peer reviewedAug 18

Perturb-ME: Scalable mechanism discovery from phenotype-enriched genome-wide screens

Perturb-ME combines phenotype enrichment, genome-scale CRISPR, and multimodal single-cell readouts to map mechanisms behind a selected phenotype. Phenotype enrichment can concentrate sequencing on informative cells and may make genome-scale perturbation mechanisms more tractable at single-cell resolution. The demonstration centers on one phenotype in a melanoma-cell context.

First author: Hanchen WangLast author: Aviv Regev

Affiliations: Genentech, Stanford University, Broad Institute of MIT and Harvard

Source: bioRxiv preprint

More: Perturb-ME code

Preprint
Single-cell & spatial
AI agents
Data & benchmarks

Preprint · not peer reviewedAug 11

Virtual-cell verification enables self-auditing AI discovery for immune rejuvenation

A preprint combines an immune-aging score, a tokenized virtual-cell model, and an agent that revises its own compound-scoring objective. The work proposes a concrete way to make an AI discovery loop check biological phenotype, predicted perturbation, and its own optimization rule against experiments. Claims of self-auditing and immune rejuvenation are the authors' framing, not evidence of therapeutic efficacy.

First author: Yue YouLast author: Luyi Tian

Affiliations: Walter and Eliza Hall Institute of Medical Research, Westlake University, Guangzhou National Laboratory, Rectified Flow Capital Limited

Source: bioRxiv preprint

Preprint
Foundation models
Single-cell & spatial
AI agents

Preprint · not peer reviewedMay 15

S2F-Agent: Harnessing sequence-to-function models for verifiable genome interpretation

S2F-Agent coordinates genomic prediction tools with human oversight and explicit execution checks, reporting better task completion than general-purpose language models on 54 workflow queries. The framework could make specialized sequence-to-function models easier to combine and audit for variant interpretation. The benchmark does not validate diagnoses or clinical decisions from an individual's genome.

First author: Li, J.Last author: Bao, Z.

Affiliations: Zhejiang University

Source: bioRxiv / medRxiv

bioinformatics

Preprint · not peer reviewedFeb 23

In Search of Ethical Procedures for LLM-Assisted Systematic Review Production: A Proof-of-Concept Evaluation of Selected Review Components

A small preprint evaluation found that language-model screening decisions changed sharply when inclusion criteria were rewritten, while generated reviews still contained misattributed citations. For teams using AI to synthesize biomedical evidence, the work highlights the need to audit both the instructions and the cited claims. The blinded review-quality comparison involved only six specialists, and the findings do not validate automated systematic reviews.

First author: McLaughlin, L.Last author: Arries, C.

Affiliations: University of Minnesota

Source: bioRxiv / medRxiv

health informatics

Preprint · not peer reviewedJul 16

Framed RSA: Representational comparisons that honor both geometry and population-mean response preferences

Framed RSA extends representational similarity analysis so model comparisons retain information about average neural responses as well as the geometry of activity patterns. Tests using brain recordings and neural-network layers suggest that preserving this information can improve the identification of the right representational model. The comparisons cover selected neural datasets and models, not a universal test of biological mechanism.

First author: Taylor, J. E.Last author: Kriegeskorte, N.

Affiliations: Columbia University

Source: bioRxiv / medRxiv

neuroscience